You are here

Roberto Ribeiro defends master dissertation entitled "Portability and performance in heterogeneous many-core systems"

Roberto Carlos Sá Ribeiro has requested the defence of his dissertation, entitled "Portability and performance in heterogeneous many-core systems". The defence will be held in room A1 of the Departamento de Informática, on December 15, 2011, 02:00pm. The public is invited.

The examining committee is the following:

  • João M. Fernandes (UMinho) - president
  • João M. P. Cardoso (FEUP) - opponent
  • Luís P. Santos (UMinho) - supervisor

Abstract

Current computing systems have a multiplicity of computational resources with different architectures, such as multi-core CPUs and GPUs. These platforms are known as heterogeneous many-core systems (HMS) and as computational resources evolve they are offering more parallelism, as well as becoming more heterogeneous. Exploring these devices requires the programmer to be aware of the multiplicity of associated architectures, computing models and development framework. Portability issues, disjoint memory address spaces, work distribution and irregular workload patterns are major examples that need to be tackled in order to efficiently explore the computational resources of an HMS.

This dissertation goal is to design and evaluate a base architecture that enables the identification and preliminary evaluation of the potential bottlenecks and limitations of a runtime system that addresses HMS. It proposes a runtime system that eases the programmer burden of handling all the devices available in a heterogeneous system. The runtime provides a programming and execution model with a unified address space managed by a data management system. An API is proposed in order to enable the programmer to express applications and data in an intuitive way. Four different scheduling approaches are evaluated that combine different data partitioning mechanisms with different work assignment policies and a performance model is used to provide some performance insights to the scheduler.

The runtime efficiency was evaluated with three different applications - matrix multiplication, image convolution and n-body Barnes-Hut simulation - running in multicore CPUs and GPUs.

In terms of productivity the results look promising, however, combining scheduling and data partitioning revealed some inefficiencies that compromise load balancing and needs to be revised, as well as the data management system that plays a crucial role in such systems. Performance model driven decisions were also evaluated which revealed that the accuracy of a performance model is also a compromising component.

02.12.2011

home contacts RSS Feed last update: 16-Apr-2020 share facebook
Drupal theme by pixeljets.com D7 ver.1.1